Noema 2B
A 2-billion-parameter model that holds its own against far larger open models — and stays accurate at half the token budget, on your own hardware.
Noema 2B is our first release: a compact, instruction-tuned language model built on Qwen3.5-2B and post-trained to be sharper and dramatically more token-efficient. It runs entirely on your own hardware in non-thinking mode, and it is engineered for the trade every local model has to make — quality against footprint. It broadly outperforms its base and, on held-out verifiable prompts, matches or beats an open model more than four times its size.
- Parameters
- 2.0B
- Base model
- Qwen3.5-2B
- Architecture
- Hybrid — Gated DeltaNet + attention
- Context length
- 262K tokens
- Modes
- Non-thinking · thinking
- Formats
- SafeTensors · GGUF
- Precision
- F16 · Q8_0 · Q6_K · Q4_K_M
- License
- Apache 2.0
Improvements
Where Noema 2B improves on Qwen3.5-2B, its base
Math at a deployment budget
66.2 vs 55.4 for the base at a 2,048-token budget — the setting small models actually run in.
Grade-school math
+3.3 single-shot and +7.5 with self-consistency; pass@8 reaches 96.7%.
Instruction following
Prompt-strict 73 vs 65, with gains across all four IFEval metrics — and 22% fewer tokens per response.
Code that runs
HumanEval 53.7 vs 39.0 and HumanEval+ 50.0 vs 36.0 — markedly more correct, runnable code than the base.
Fewer tokens per answer
394 vs 565 tokens per correct answer on IFEval; wrong answers also fail shorter instead of running to the cap.
Reliable generation
Runaway, repetitive generation is effectively gone — failures terminate with clean, parseable answers.
Benchmarks
Higher is better · non-thinking mode
Punches above its weight
On 130 held-out verifiable prompts never seen in training, Noema 2B matches or beats Qwen3.5-9B — an open model over four times its size — on 84% of items.
Budget-robust
Share of full-budget MATH-500 accuracy retained at half the token budget. Noema stays near its ceiling; the base loses nearly a fifth of its performance when tokens are tight.
Versus the base — and a 9B reference
GSM8K
k=1GSM8K
SC@8MATH
100-setMBPP-100
pass@1IFEval
prompt-strictAll three scored on the same frozen harness (non-thinking) — identical eval for each. Qwen3.5-9B, 4.5× Noema's size, is a reference ceiling: Noema 2B closes much of the base → 9B gap.
Coding — versus the base
HumanEval
pass@1HumanEval+
pass@1MBPP
pass@1MBPP+
pass@1Noema 2B vs Qwen3.5-2B on our frozen harness — non-thinking, pass@1. EvalPlus suite: HumanEval / MBPP plus their extended-test (+) variants. (Current general models publish LiveCodeBench, not EvalPlus, so there is no fair current-peer set here.)
Versus the current 2–4B class
IFEval
GSM8K
MATH-500
Non-thinking mode for all. LFM2-2.6B from the LFM2 technical report (arXiv:2511.23404; no reasoning mode). SmolLM3-3B from community /no_think evals — GSM8K flexible-extract, IFEval prompt-level loose (strict not published); MATH-500 pending. Qwen3.5-4B non-thinking pending. Noema on our frozen harness (MATH-500 at a 2,048-token budget). Protocols differ across labs.